🐶 Pix2Pix GAN for Image Generation from Edge Maps

This project implements a Pix2Pix Conditional GAN that generates realistic images from edge outlines, using object masks from the COCO 2017 dataset. The model learns to translate binary edge maps into colored, high-resolution dog/object images using a U-Net-based generator and a PatchGAN discriminator.

---

🧠 Project Summary

• Dataset: COCO 2017 via TensorFlow Datasets (TFDS)
• Input: Edge maps (generated using OpenCV's Canny filter on segmentation masks)
• Output: 256×256 RGB images of objects (e.g., dogs)
• Model: Pix2Pix GAN with U-Net generator + PatchGAN discriminator
• Loss Function: GAN loss + L1 loss (total = GAN + λ * L1)
• Frameworks: TensorFlow 2.x, OpenCV, Matplotlib

---

🚀 Highlights

• 🧩 End-to-end edge-to-photo translation using paired image data
• 🖼️ Real-time image generation visualization
• 📦 Custom training loop with TensorBoard support
• 🧪 Includes data preprocessing, augmentation, batching, and shuffling
• 💾 Checkpoint saving for long training runs

---

📂 Directory Structure

📁 pix2pix-gan-coco/
│
├── ai_project_ares.py         # Full training & model definition
├── checkpoints/               # Saved model weights
├── logs/                      # TensorBoard logs
├── samples/                   # Example predictions
├── README.md                  # You are here

---

🔧 How It Works

1. Data Preprocessing:

   • Loads COCO 2017 images using TFDS
   • Extracts segmentation masks and applies edge detection (Canny)
   • Resizes and normalizes both edge maps and original images

2. Model Architecture:

   • Generator: U-Net-style encoder-decoder with skip connections
   • Discriminator: PatchGAN to classify real/fake (image, outline) pairs

3. Training Loop:

   • Uses tf.GradientTape for manual backpropagation
   • Logs losses (GAN, L1, Discriminator) with TensorBoard
   • Saves model checkpoints every 5000 steps
   • Generates preview outputs during training

---

📊 Sample Outputs

Input Edge Map	Ground Truth	Generated Image

---

🛠️ Tech Stack

• Python 3.10+
• TensorFlow 2.14+
• OpenCV
• Matplotlib
• COCO Dataset (via tensorflow_datasets)

---

📈 To Train the Model

# Clone the repo
git clone https://github.com/your-username/pix2pix-gan-coco.git
cd pix2pix-gan-coco

# Install dependencies
pip install tensorflow opencv-python matplotlib tensorflow-datasets

# Run training script
python ai_project_ares.py

---

📌 Future Improvements

• Add FID and SSIM metrics for evaluation
• Support for custom datasets (e.g., Stanford Dogs)
• Streamlit or Gradio UI for demo
• Export model for real-time inference

---

📄 License

This project is released under the MIT License.

---

🙋‍♂️ Author

Anurag Pradhan
📧 anuragpradhancb@gmail.com
🔗 LinkedIn • GitHub

---